Amazon Threat Intelligence disclosed on November 20 that China-linked hackers directed agents through the AI penetration framework Cairn to launch attacks on medical imaging servers at multiple East Asian universities in September. The agent discovered a path traversal zero-day vulnerability in the open-source project MONAI Label and escalated from API mapping to root control within 11 minutes.
Reconstruction of Events
The investigation described by Amazon CISO CJ Moses began on September 10, when a publicly accessible server was found hosting a Cairn instance, allowing its operation history and prompts to be viewed. The prompts were in Simplified Chinese, and timestamps were set to China Standard Time. The targets were East Asian universities. Cairn was instructed to “gain shell access” to an IP address. One attempt lasted 16 hours and executed 148 actions without success, including a failed Log4Shell attempt; another attempt tested more than 30,000 credential pairs over more than six hours and also failed. The only successful case was an agent discovering and exploiting the zero-day path traversal vulnerability in a medical imaging AI application, which was then used to compromise a second university. Amazon Threat Intelligence assessed it as a zero-day because there was no corresponding record in public databases. The maintainers released a patch on November 18, and AWS notified relevant organizations together with government partners.
Mechanism Breakdown
Cairn allows operators to provide a starting point and objective in natural language, with multiple AI agents coordinating planning, execution, and path adjustment based on feedback, without requiring humans to specify each step. The attacker only input an IP and the instruction “gain shell access,” and the agent autonomously completed mapping, vulnerability search, and exploitation. The same framework was previously recorded by Gambit Security in an intrusion against an online retailer, and Google researchers also observed similar use in an intrusion at a U.S. aerospace company. Cairn and tools such as Artex are open-source autonomous multi-agent offensive frameworks; the difference is that Cairn lacks internal structural constraints specific to cybersecurity.
Industry Impact
This incident shows that AI agents can independently complete the full “vulnerability discovery → exploitation” chain in real attacks, at a speed exceeding the time most security operations centers need to manually triage alerts. As an open-source medical imaging framework, MONAI Label’s patch was released after the attack occurred, highlighting the exposure window of open-source components under AI-assisted attacks. Several cloud vendors and research institutions have successively issued warnings about frameworks such as Artex and Cairn, indicating that both attackers and defenders are paying increasing attention to autonomous AI tools.
Strategic Assessment
(The following is analysis, not fact.) Current security evaluation frameworks mostly focus on defensive scenarios and inadequately cover “offensive autonomous vulnerability discovery” capabilities; testing dimensions may need to be expanded to include general-purpose agent behavior patterns like Cairn’s. Because open-source tools lack built-in cybersecurity structures, their use cases are harder to predict, and in the future use-boundary restrictions may need to be added at the framework design stage to reduce the risk of abuse.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接